Neural Network Based Pattern Recognition in Visual Inspection System for Intergrated Circuit Mark Inspection

Industrial visual machine inspection system uses template or feature matching methods to locate or inspect parts or pattern on parts. These algorithms could not compensate for the change or variation on the inspected parts dynamically. Such problem was faced by a multinational semiconductor manuf...

Full description

Bibliographic Details
Main Author: Sevamalai, Venantius Kumar
Format: Thesis
Language:English
English
Published: 1998
Subjects:
Online Access:http://psasir.upm.edu.my/id/eprint/10131/1/FK_1998_8_A.pdf
Description
Summary:Industrial visual machine inspection system uses template or feature matching methods to locate or inspect parts or pattern on parts. These algorithms could not compensate for the change or variation on the inspected parts dynamically. Such problem was faced by a multinational semiconductor manufacturer. Therefore a study was conducted to introduce a new algorithm to inspect integrated circuit package markings. The main intend of the system was to verify if the marking can be read by humans. Algorithms that the current process uses however, was not capable in handling mark variations that was introduced by the marking process. A neural network based pattern recognition system was implemented and tested on images resembling the parts variations. Feature extraction was made simple by sectioning the region of interest (ROI) on the image into a specified (by the user) number of sections. The ratio of object pixels to the entire area of each section is calculated and used as an input into a feedforward neural network. Error-back propagation algorithm was used to train the network. The objective was to test the robustness of the network in handling pattern variations as well as the feasibility of implementing it on the production floor in tetms of execution speed. Two separate programme modules were written in C++; one for feature extraction and another for neural networks classifier. The feature extraction module was tested for its speed using various ROI sizes. The time taken for processing was round to be almost linearly related to the ROJ size and not at all effected by the number of sections. The minimum ROJ setting (200 X 200 pixels) was considerably slower at 5 5ms compared to what was required - 20ms. The neural networks c1assifier was very successful in classifying 1 3 different image patterns by learning from 4 training patterns. The classifier also clocked an average speed of 9.6ms which makes it feasible to implement it on the production floor. As a final say, it can be concluded that by carefully surveying the choices of hardware and software and its appropriate combination, this system can be seriously considered for implementation on the semiconductor production floor.